An identification apparatus performs classification using a plurality of classifiers, and calculates the reliability of its classification result. A data obtaining unit obtains input data. A feature quantity obtaining unit obtains a feature quantity corresponding to the input data. A plurality of classifiers receive input of the feature quantity and perform classification based on the input feature quantity. An identification unit inputs the feature quantity into each of the classifiers, and generates a single second classification result based on a plurality of classification results obtained from the classifiers. A reliability generation unit generates a reliability of the second classification result based on variations across the plurality of classification results.
Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. An identification apparatus, comprising: data obtaining circuitry that obtains input data; feature quantity obtaining circuitry that obtains a feature quantity corresponding to the input data; a plurality of classifiers that receives input of the feature quantity, performs classification based on the input feature quantity, and outputs a single first class value, which is a value corresponding to a class obtained by the classification, respectively; identification circuitry that inputs the feature quantity into each of the classifiers, and generates a second class value, which is a single classification result, based on a plurality of the first class values obtained from the classifiers; and reliability generation circuitry that generates a reliability of the second class value based on variations across the plurality of the first class values, wherein the reliability generation circuitry generates the reliability so that the magnitude of the variation of the plurality of the first class values and the magnitude of the reliability of the second class value have a negative correlation.
An identification system analyzes input data using multiple classifiers to improve accuracy and assess confidence. It obtains input data and extracts relevant features. These features are fed into several classifiers, each outputting a class prediction. The system combines these individual predictions into a single, final class prediction. Crucially, it also calculates a reliability score for this final prediction based on the agreement (or disagreement) among the individual classifier outputs. Higher variation in individual classifier results leads to lower reliability in the final classification. This relationship is inversely proportional.
2. The identification apparatus according to claim 1 , wherein the reliability generation circuitry generates the reliability of the second class value based on a variance or a standard deviation of class values output from the classifiers.
The identification system described previously calculates the reliability of its classification by analyzing the spread of class values outputted by the multiple classifiers. Specifically, the system uses the variance or standard deviation of these class values as a measure of reliability. A higher variance/standard deviation (more disagreement) indicates lower reliability, while a lower variance/standard deviation (more agreement) indicates higher reliability.
3. The identification apparatus according to claim 2 , wherein the reliability generation circuitry determines the variance or the standard deviation using a median or a mode of the class values output from the classifiers.
Building on the identification system that calculates reliability using variance or standard deviation of classifier outputs, this version refines that calculation. It uses the median or mode of the individual classifier outputs as a central point to calculate the variance or standard deviation. This helps to reduce the influence of outliers or biased classifiers when determining the overall reliability score of the final classification result.
4. The identification apparatus according to claim 1 , further comprising, evaluation circuitry that evaluates an accuracy of each of the classifiers, wherein the identification circuitry or the reliability generation circuitry weights the plurality of the first class values in accordance with the evaluated accuracy of the corresponding classifier, and generates the second class value or the reliability.
The identification system described earlier is enhanced by incorporating classifier accuracy. Before combining the individual classifier predictions and calculating reliability, the system evaluates the accuracy of each individual classifier. The final class prediction or the reliability score is then weighted based on these accuracy scores. More accurate classifiers have a greater influence on the final result and reliability score, while less accurate classifiers have less influence.
5. The identification apparatus according to claim 4 , wherein the evaluation circuitry evaluates the accuracy of each classifier using test data.
Expanding on the identification system that evaluates classifier accuracy, this implementation details how that accuracy is measured. The system uses test data, separate from the training data, to evaluate the performance of each classifier. By running the test data through each classifier and comparing the results to known correct answers, the system determines an accuracy score for each classifier. This accuracy score is then used to weight the classifier's contribution to the final result, as previously described.
6. The identification apparatus according to claim 4 , wherein the evaluation circuitry evaluates the accuracy of each classifier using the number of learning samples used for learning of the classifier.
Instead of using test data to evaluate classifier accuracy, this identification system variant uses the number of learning samples used to train each classifier as a proxy for accuracy. Classifiers trained on more data are assumed to be more accurate. The system uses the number of training samples to weigh the individual classifier outputs when generating the final class value or when assessing the reliability of the classification result.
7. The identification apparatus according to claim 1 , wherein the input data comprises an image.
In the identification system described earlier, the input data being classified is an image. This could be an image of a face, an object, or any other visual data that the system is designed to analyze and categorize. The image is preprocessed to extract relevant features, which are then fed into the multiple classifiers.
8. The identification apparatus according to claim 7 , wherein a target for classification performed by each classifier comprises at least one of an attribute or a state of a person included in the image.
In the image-based identification system, the classifiers are used to identify attributes or states of people within the image. For example, the classifiers might identify a person's age range, gender, emotional state (happy, sad, angry), or clothing type. Each classifier is responsible for identifying a specific attribute or state. The final output could be a combination of these identified attributes, providing a more complete description of the person in the image.
9. A method for controlling an identification apparatus comprising a plurality of classifiers configured to perform classification based on an input feature quantity, the method comprising: obtaining input data; obtaining a feature quantity corresponding to the input data; inputting the obtained feature quantity into each of the classifiers, which outputs a single first class value, which is a value corresponding to a class obtained by the classification, respectively; generating a second class value, which is a single classification result, based on a plurality of the first class values obtained from the classifiers; and generating a reliability of the second class value based on variations across the plurality of the first class values, wherein the reliability is generated so that the magnitude of the variation of the plurality of the first class values and the magnitude of the reliability of the second class value have a negative correlation.
A method for identification using multiple classifiers involves these steps: First, input data is acquired. Next, relevant features are extracted from the input data. These features are input into multiple classifiers, each of which outputs a single class prediction. Then, a final class prediction is generated based on the individual classifier outputs. Finally, a reliability score is calculated for the final prediction based on the agreement or disagreement among the individual classifier outputs. High variation among classifiers reduces the reliability score, which is inversely proportional.
10. A non-transitory computer readable storage medium recording a computer program for causing a computer to perform a method comprising the steps of: obtaining input data; obtaining a feature quantity corresponding to the input data; inputting the obtained feature quantity into each of the classifiers, which outputs a single first class value, which is a value corresponding to a class obtained by the classification, respectively; generating a second class value, which is a single classification result, based on a plurality of the first class values obtained from the classifiers; and generating a reliability of the second class value based on variations across the plurality of the first class values, wherein the reliability is generated so that the magnitude of the variation of the plurality of the first class values and the magnitude of the reliability of the second class value have a negative correlation.
This claim describes a computer program stored on a non-transitory medium that, when executed, performs an identification method. The method involves: acquiring input data; extracting relevant features; inputting these features into multiple classifiers, each outputting a class prediction; generating a final class prediction based on the individual outputs; and calculating a reliability score for the final prediction based on variations between the classifiers' predictions. High variation among individual predictions reduces the final reliability, which has an inversely proportional relationship.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
August 31, 2015
June 27, 2017
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.